Download The Role of Social and Cognitive Abilities in the Emergence of

The Role of Social and Cognitive Abilities in the Emergence of
Communication: Experiments in Evolutionary Robotics
Davide Marocco1,3
University of Calabria,
Centro Interdipartimentale della
Comunicazione
Arcavacata di Rende, 87036
Cosenza, Italy
davidem@ip.rm.cnr.it
1
Angelo Cangelosi2
University of Plymouth
Institute of Neuroscience
and School of Computing
Drake Circus, PL4 8AA
Plymouth, UK
acangelosi@plymouth.ac.uk
2
Abstract
Evolutionary robotics is a biologically inspired approach to
robotics that is advantageous to studying the evolution of
language. A new model for the evolution of language is
presented. This model is used to investigate the
interrelationships between communication abilities, namely
linguistic production and comprehension, and other
behavioral skills. For example, the model supports the
hypothesis that the ability to form categories from direct
interaction with an environment constitutes the ground for
subsequent evolution of communication and language. A
variety of experiments, based on the role of social and
evolutionary variables in the emergence of communication,
are described.
1. Introduction
The communication between autonomous agents, be they
robots or simulated virtual agents, has recently attracted the
interest of researchers from different fields. In engineering,
the design and evaluation of communication systems is
interesting due to its practical applications for agent-agent
interaction and also for human-agent and human-robot
communication (e.g. Lauria et al., 2002). For cognitive
scientists, the development of computational models for the
evolution of language permits the investigation of the role of
sensorimotor, cognitive, neural and social factors in the
emergence and establishment of communication and
language (Cangelosi & Parisi, 2002).
Studies on the emergence of communication are often
based on synthetic methodologies such as adaptive behavior
and artificial life (Steels, 1997; Kirby, in press). A group of
autonomous agents interact via language games to exchange
information about the external environment. Their
coordinated communication system is not externally
imposed by the researcher, but emerges from the interaction
between agents. In such models, the levels of detail of the
representation of the agents and of their environment can
vary significantly. This constitutes a continuum between
abstract point models, at one end, and situated, embodied
robots at the other. At one extreme, only the essential
communicative properties of the agents and the environment
are simulated. For example, the environment can consist of
Stefano Nolfi3
National Research Council
Institute of Cognitive Science
and Technologies,
Viale Marx 15, 00137
Rome, Italy
nolfi@ip.rm.cnr.it
3
a list of abstract “meanings”, and the agent consists of a
function, or rule set, that maps these meanings to signals
(e.g. Kirby, 2001; Oliphant, 1999). This approach is useful
when one wants to study the dynamics of the autoorganization of lexicons and syntax and its dependence on
single, pre-identified factors. An intermediate approach to
language evolution is based on grounded simulation models
(Harnad, 1990). The agents’ environment is modeled with a
high degree of detail upon which emergent meanings can be
directly grounded. Each simulated agent has a set of
sensorimotor, cognitive and social abilities that allow it to
build, through interaction, a functional representation of the
environment and use it to communicate (e.g. Cangelosi,
2001; Cangelosi & Harnad, 2000; Hazlehurst & Hutchins,
1998). This type of models supports the investigation of the
interaction amongst various abilities of the agents for the
emergence of language and the grounding of
communication symbols in the environment and the agent’s
behavior.
At the other end of the continuum, the communicative
behavior of embodied and situated robots results from the
dynamical interaction between its physical body, the
nervous and cognitive system and the external physical and
social environment (Beer, 1995). For example, robots can
interact and communicate among themselves (e.g. Steels &
Vogt, 1997; Quinn, 2001), with virtual Internet agents
(Steels, 1999) and with humans (Steels & Kaplan, 2000).
Such an approach permits the study of the interaction
between the different levels of a behavioral system, that is
from sensorimotor coordination to high-level cognition and
social interaction.
Amongst the robotic approaches to studying adaptive
behavior, evolutionary robotics (Nolfi & Floreano, 2002)
offers a series of advantages. Through evolutionary
experiments, artificial organisms autonomously develop
their behavior in close interaction with their environment.
The main advantages of this approach are: (a) it involves
systems that are embodied and situated (Brooks, 1991;
Pfeifer and Scheier, 1999), and (b) it is an ideal framework
for synthesizing robots whose behavior emerge from a large
number of interactions among their constituent parts. This
can be explained by considering that, in evolutionary
1
experiments, robots are synthesized through a selforganization process based on random variation and
selective reproduction where the selection process is based
on the behaviors that emerge from the interactions among
the robot's constituent elements and between these elements
and the environment. This allows the evolutionary process
to freely exploit interactions without the need to understand
in advance the relation between interactions and emerging
properties as it is necessarily required in other approaches
that rely more on explicit design.
For these reasons the evolutionary robotics approach has
been successfully applied to the synthesis of robots able to
exploit sensorimotor coordination (Nolfi, 2002); on-line
adaptation (Nolfi and Floreano, 1999); body and brain coevolution (Lipson and Pollack, 2000); competing and
cooperative collective behaviors; (Nolfi and Floreano, 1998,
Martinoli, 1999; Baldassarre, Nolfi, and Parisi, 2002).
These advantageous aspects of evolutionary robotics are
of particular importance for modeling the evolution of
language and communication. Sensorimotor coordination,
social interaction, evolutionary dynamics and the use of
neural systems can all have a potential impact in the
emergence of coordinated communication. In this paper,
new experiments are presented that study the emergence of
communication in evolutionary robotics models. They are
based on recent work by Nolfi and Marocco (2002) for the
emergence of sensorimotor categorization. Nolfi and
Marocco evolved the control system of artificial agents that
are asked to categorize objects with different shapes on the
basis of tactile information. Each agents uses proprioceptive
information to actively explore objects using a threesegment arm. In addition, the agent uses the activation of
one output node of its neural network controller as input.
Agents are selected only for their performance in
discriminating (categorizing) the objects using this unit, not
for their ability to explore them. This results in the
emergence of an active tactile exploration strategy that
differentiate between objects of different shapes. Nolfi and
Marocco’s model is an example of explicit selfcategorization.
In this new model, the robotic agents share the explicit
categorization of objects. That is, the activation of the
output nodes is the signal (“name”) sent to another agent to
instruct it on what to do with the object. Agents will be
selected on their ability to manipulate objects correctly, not
on their (linguistic) ability to name them correctly. A variety
of experiments will test the role of different social and
evolutionary factors. Results will be used to analyze the role
of sensorimotor, social and cognitive factors in the
emergence of communication. The direct relations between
behavioral and communication abilities, such us language
production and comprehension, will also be discussed.
2. Method
The behavior of each agent consists of exploration within
the environment, on the basis of tactile information, and the
communication, about the type of objects that are in it. The
environment consists of an open three-dimensional space in
which one of two different objects is present in each epoch
(Figure 1). The two objects used in this simulation are a
sphere and a cube.
Figure 1 – The arm and a spherical object.
Figure 2 – A schematic representation of the arm.
Agents are provided with a 3-segments arm with 6
degrees of freedom (DOF) and extremely course touch
sensors (see Figure 2). Each segment consists in a basic
structure of two cylindrical bodies and two joints. This is
replicated for three times. The basic structure consists of a
shorter body of radius 2.5 and length 3 and a longer body of
the same radius and length 10 for the first two segments.
The length of the third segment is 5. This shorter segment
represents a fingerless hand. The two bodies of each
segment are connected by means of a joint (i.e. the Joint E
in the Figure) that allows only one DOF on axis Y, while the
shorter body is connected at the floor, or at the longer body,
by means of a joint (i.e. the Joint R) that provides one DOF
on axis X. In practice, the Joint E allows to elevate and
lower the connected segments and the Joint R allows to
rotate them in both direction. Notice that Joint E is free to
moves only in a range between 0 and π/2, just like a human
arm that can bend the elbow solely in a direction. The range
2
of Joint R is [–π/2, +π/2]. Gravity is {0, –1, 0}. Each
actuator is provided with a corresponding motor that can
apply a maximum force of 50. Therefore, to reach every
position in the environment the control system has to
appropriately control several joints and to deal with the
constraints due to gravity.
The sensory system consists of a simple contact sensor
placed on each longer body that detects when this body
collides with another, and two proprioceptive sensors that
provide the current position of each joint.
The controller of each individual consists of an artificial
neural networks with 11 sensory neurons connected to 3
hidden neurons. These connect with 8 output neurons. The
first 9 sensory neurons encode the angular position
(normalized between 0.0 and 1.0) of the 6 DOF of the joints
and the state of the three contact sensors located in the three
corresponding segments of the arm. The other 2 sensory
neurons receive their input from the other agents. The first 6
motor neurons control the actuators of the corresponding
joints. The output of the neurons is normalized between [0,
+π/2] and [–π/2, +π/2] in the case of elevation or rotational
joints respectively and is used to encode the desired position
of the corresponding joint. The motor is activated so to
apply a force (up to 50) proportional to the difference
between the current and the desired position of the joint.
The last 2 output neurons encode the signal to be
communicated to the other agents. This works as a small
winner-takes-all cluster, where the neuron with the highest
activation is set to 1 and the other to 0.
The activation state of internal neurons was updated
accordingly to the following equations (output neurons were
updated according to the logistic function):
j
O j = τ jOj
= tj +
( t −1)
wijOi
(
+ (1 − τ j ) 1 + e
)
− Aj −1
(1)
0 ≤τ j ≤1
With Aj being the activity of the jth neuron (or the state
of the corresponding sensor in the case of sensory neurons),
tj the bias of the jth neuron, Wij the weight from the ith to
the jth neuron, Oi the output of the ith neuron. Oj is the
output of the jth neuron, τj the time constant of the jth
neuron.
Each individual was tested for 36 epochs, each epoch
consisting of 150 sensorimotor cycles. At the beginning of
each epoch the arm is fully extended. A spherical or a cubic
object is placed in a random selected position in front of the
arm. The position of the object is randomly selected
between the following intervals: 15.0 <= X <= 25.0; Y =
7.5; –5.0 <= Z <= 5.0). The object is a sphere (15 units in
diameter) during even epochs and a cube (15 units in side)
during odd epochs so that each individual has to
discriminate the same number of spherical and cubic objects
during its lifetime.
In addition to the proprioceptive information, agents
also receive in input a 2-bit signal produced by some other
agent in the population, such as the parent or any agent from
the population (linguistic comprehension task). The protocol
of interaction and communication between agents was
systematically varied and is analyzed in section 3.
Before they act as speaker, agents undergo a linguistic
production task. That is, each agent is put in the
environment and asked to interact with the object. The value
of the two output neurons in the last cycle of the epoch is
saved and used as the signal produced to “name” the object.
A genetic algorithm is used to evolve the behavior of
agents. The genotype of each agent consists of 81
parameters that include 67 weights, 11 biases, and 3 time
constants. Each parameter is encoded with 8 bits. Weights
and biases are normalized between –5.0 and 5.0, time
constants are normalized between 0.0 and 1.0.
The fitness rewards the behavior of the agent with the
current object in the environment. Good communication
behavior does not produce any fitness gain for the speaker.
Following the behaviors evolved in Nolfi & Marocco’s
(2002) simulation, the agent has to touch and stay in contact
with one object (the sphere) and has to avoid as much as
possible to touch the other object (cube). The fitness of
individuals is computed by summing the number of cycles
in which the agent touches the sphere or does not touch the
cube. Fitness scores decrease for each cycle the agent
touches the cube or when it does not touch the sphere.
A population of 80 agents is used in each simulation.
During selection, the 20 agents with the highest fitness (i.e.
behavioral performance) reproduce and each make 4
offspring. The genotype of each offspring is then subject to
mutation with an overall probability of 2%. That is, each bit
has a 2% probability of being mutated, by generating a
random binary value. There is generational overlap between
the population of parents and that of new offspring. The first
will only act as speakers and cannot reproduce anymore.
The population of new offspring will be subject to the
fitness test and will reproduce at the end of their lifetime.
Evolutionary simulation of embodied robotic agents can
be time consuming and computationally expensive. To
reduce the time necessary to test individual behaviors and to
model the real physical dynamics as accurately as possible,
the rigid body dynamics simulation SDK of VortexTM was
used1. This was linked to the EvoRobot simulator (Nolfi,
2000).
3. Results
The simulation model was used to run a series of
experiments on the role of various social and evolutionary
variables in the emergence of shared communication. The
first independent variable refers to the selection of speakers
(SPEAKER) with two levels: Parent or All. In the first case,
each agent receives communication signals only from its
own parent. In the second level of the variable, each agent
1
http://www.cm-labs.com/products/vortex/
3
can receive signals from any individual of the previous
population. This factor is aimed at investigating the role of
different social groups of speakers in facilitating shared
communication.
The second independent variable manipulated during
experiments consists in the time in which communication is
allowed (COMMUNICATION) with two levels: From_0
and From_50. In the first case, agents were allowed to
communicate from the initial random generation. In the
second level of the variable, agents start to communicate
between themselves only at generation 50, i.e. after they
have evolved a good ability to touch/avoid the two objects.
Through this variable it will be possible to investigate the
initial behavioral and cognitive abilities necessary to evolve
communication.
For each of the 4 conditions (2 SPEAKER × 2
COMMUNICATION), 10 replications were executed, by
changing the initial random population. Fifty generations
were necessary to pre-evolve an optimal behavior of object
manipulation to be used in the From_50 conditions. Table 1
reports the communication success in each condition in
terms of good populations and percentage of good speaker
in the population. The criterion for deciding whether a
population has successfully evolved communication
depends on the fact that, at the last generation, at least 50%
of agents produce two signals that differentiate the two
objects.
When agents listen to all individuals of the previous
generation, no stable communication exists in the last
generations. In fact, during evolution good lexicons
sometimes emerge for a short time, but they are not
maintained or further developed by the whole population. A
temporary good lexicon is defined as the case in which at
least 20% of agents use two different signals to name the
two objects. In 8 of the 10 From_50 - All speaker
populations, such temporary appearances of good signal
production is observed. Figure 4 shows the best population
in the From_50 - All speaker conditions. Here the longest
period of good production only lasts for 17 generations,
with a maximum peak of best language at 41%.
Table 1 – Data on the emergence of communication in each
experimental condition. The first line contains the number of
populations (out of 10) where communication emerged. The
second line contains the average percentage of good speakers
for the 10 replications and the average for the best performing
population (value between brackets).
Figure 3 – Data for the best population of the condition Parent
speaker - From_50.
1
Fitness(all)
N. speakers
0.8
0.6
0.4
0.2
0
1
51
101
151
-0.2
1
Fitness(all)
N. speakers
0.8
SPEAKER
Parent
# good pops
% speakers (best pop)
All
# good pops
% speakers (best pop)
COMMUNICATION
From 0
COMMUNICATION
From 50
5
7
27% (75%)
63% (100%)
0
0
7% (20%)
5% (27%)
The results of the number of populations that evolve
shared communication clearly show that it is only when the
parents act as the speakers there is a selective pressure for
the emergence and preservation of a shared communication
system. In particular, 7 populations out of 10 reach a stable
communication system when language is introduced after
agents have learned to use both objects. Figure 3 shows the
fitness curves and the proportion of good speaker in the best
seed of the condition From_50 - Parent speaker.
When communication is introduced directly from the
initial random population, the probability of evolving a good
language, together with a good behavior, is lower (5
populations out of 10). This advantage for evolving
languages after the basic behavioral skills have evolved is
similar to that observed by Cangelosi & Parisi (2001) in a
grounded simulation model on the emergence of verbs and
nouns.
0.6
0.4
0.2
0
1
51
101
151
-0.2
Figure 4 – Data for the best population in condition All speakers
- From_50.
The lexicon produced by agents in successful
replications has been tested to investigate whether
individuals actually use this language in a meaningful way,
i.e. avoid the cube when the signal produced in response to
the cube is used, and touch the sphere when the other signal
is used. Figure 5 shows the behavior of an agent that
interacts with the cube with or without language. This tests
the linguistic comprehension ability of agents. The pictures
on the left column (Figure 5 - left) show the behavior of the
agent when no input signal is used. The agent needs to touch
the cube, at least once (in cycle 95), to identify it as a cube
4
and then retract from it. The pictures on the right (Figure 5 right) show the behavior of the agent when the signal “10”
is used as additional input. This signal is produced by the
parent organism at the end of the interaction with a cube.
During this scene, the agent does not need to touch the cube
at all because the signal “10” identifies it as a cube. The
meaning of “10” can be interpreted as “cube”2, because the
listener treats the object as a cube, and the speaker produces
it after its interaction with a cube. When the signal “01” is
used, the agent touches the object regardless of its shape. In
this case, “01” has the meaning of “sphere”.
speakers reach the highest fitness scores, with a significant
advantage for the From_50 populations (e.g. average fitness
of best individuals = 0.55; fitness peak in best population =
0.72) versus the From_0 population (average = 0.45, peak =
0.66). The baseline for the behavior without communication
is the fitness at generation 50 of the From_50 simulation,
before agents start to communicate (average = 0.44, peak =
0.52). Consider that the maximum hypothetical fitness score
is 1. This can never be reached because, for example, at the
beginning of each epoch some negative fitness cycles are
always necessary for agents to reach the spherical object and
start gaining fitness.
4. Discussion
Figure 5 – Agent’s interaction with the cube and test of linguistic
understanding ability. Left column: Only the proprioceptive input
is given to the agent. Right column: An additional
communication signal is given as input. This is produced by
another agent at the end of its interaction with a cube. Figures
from the best individual of a From_50 - Parent speaker
population.
Fitness data shows that final scores in the 4 experimental
conditions reflect the pattern of results on the emergence of
successful communication. The two conditions with Parent
2
This signal can also be interpreted as the verb avoid, instead of as
the noun cube. In fact, in this model it is not possible to distinguish
between syntactic word classes (cf. Cangelosi & Parisi 2001 and
Cangelosi 2001 for a discussion)
There are several issues that can be discussed regarding
these results, and what we can learn from the model. For
example, the following questions can be asked: (1) Is there
any benefit for the agents to be in a population where good
communication has emerged? (2) Is there any direct
advantage to evolving a good linguistic comprehension
ability? (3) Is there any direct advantage to evolving good
linguistic production abilities? (4) What is the relation
between comprehension, production and behavioral
abilities? (5) What are the underlying factors that cause and
favor the emergence of communication?
To answer the first question, it is possible to compare the
fitness results in the simulations where no shared
communication emerged, and those where good
communication systems evolved. The condition in which
communication emerged more frequently (From_50, Parent
speaker) will be used as example. In this condition, 7
populations evolved good languages, whilst 3 did not.
Figure 6 shows the average fitness of the good
communication populations (thick lines) and that of the no
communication populations (thin lines). The chart clearly
shows that agents who use communication reach fitness
values that are higher that those not communicating. This is
true both for the fitness of the best individual and for that of
the whole population. For example, at the final generation
the average fitness of the 7 successful communication
replications is 0.35, while it is 0.21 for the 3 unsuccessful
populations. Moreover, the fitness in these 3 populations
remains relatively constant during the simulation. In the first
50 generations after communication is permitted (i.e. from
50 to 100), there is no increase and the average fitness at
generation 100 is very similar to that at generation 50. In the
remaining generations, the agents gain some extra fitness
points, which are due to the continuation of the evolutionary
algorithm search.
The extra fitness gain in populations that evolve
communication is easily explained by the direct benefits for
the behavior (i.e. fitness) of using two different signals: one
for the cube, and one for the sphere. As already shown in
Figure 5, during the interaction with a cube the input of its
“name” produces significant improvements to behavioral
performance. Agents do not need to touch the object to
5
recognize it, and therefore do not lose fitness due to such
exploratory behavior. In addition, they gain fitness in every
cycle. There is also some benefit for the use of the signal for
the sphere. If an agent initially is told that there is a
spherical object in the environment, it can go directly
towards the object and touch it, without having to use some
interaction cycles for recognizing the object as a non cube.
0.7
Fitness(best) - Communication
Fitness(all) - Communication
Fitness(best) - No communication
Fitness(all) - No communication
0.6
0.5
Fitness
0.4
0.3
0.2
0.1
0
0
25
50
75
100
125
150
-0.1
Generations
Figure 6 – Average fitnesses of the conditions From_50 Parent speaker. Thick lines refer to the average fitness of the 7
replications where good communication emerged (continuous
line for the best agent and dotted line for the average of all
agents). Thin lines refer to the average fitness of the 3
replications where no shared communication emerged.
These explanations also answer the second question,
since they identify a direct adaptive advantage for evolving
a good comprehension ability. The other question regarding
the “direct” advantage of linguistic production abilities is
more difficult to answer. In fact, there seems to be no direct
fitness advantage to the agents to speaking well. Individuals
only update their fitness when they hear others speaking.
When agents act as speakers, some have already
reproduced, whilst the others have not been selected at all.
In the condition Parent speaker, agents only speak to their
own children. Therefore, the kinship relationship can
partially explain this apparent altruistic behavior and the
indirect fitness gain for the common genes shared by the
parent and its offspring (e.g. Ackley & Littman, 1994). The
benefits of kin selection can also explain the successful
evolution of communication in the Parent speaker versus the
All speaker conditions. However, there is another important
phenomenon to be considered. In the Parent speaker
conditions, the linguistic input to each listener is constant,
since its parent will always use the same signal for the same
object. In addition, when the parent is a good speaker (i.e. it
uses two different signals to refer to the two objects), its
signals are more reliable. The child can then try to use them
to improve its fitness performance. In the All speaker
conditions, the high variability of the linguistic input
coming from all agents of previous generation can be too
unreliable, and agents will tend to ignore it.
In the All speaker conditions, some communication
abilities also emerge, although the number of good speakers
never reaches the critical amount needed to allow it to
remain stable until the end of the simulation (cf. Figure 4).
In addition, in the Parent speaker conditions, there are three
cases when shared communication does not evolve.
According to the altruistic, kin selection explanation, all
Parent speaker populations should evolve communication
because of it indirect adaptive advantage. The fact that this
does not always happen raises the issues of understanding
the relation between linguistic comprehension/production
abilities and other behavioral/cognitive abilities (question
4), and the identification of factors that cause and favor the
emergence of shared communication (question 5). First, the
data in Table 1 indicates that it is easier to evolve good
communication when language is introduced after the preevolution of good behavioral capacities (7 out of 10
populations) than when agents are allowed to communicate
from the initial generation (5 out of 10 seeds). In addition,
the onset of effective communication (i.e. when at least 20%
of agents speak well) is much earlier in the From_50
populations (on average after 16 generations) that in the
From_0 simulations (on average after 41 generations). This
data is consistent with Cangelosi and Parisi’s (2001) model
on the evolution of syntactic languages. This research
showed that agents learn languages more efficiently when
communication is introduced after the pre-evolution of good
behavioral skills. Effectively, the pre-evolution of good
behavior “prepares” a cognitive ground upon which good
linguistic abilities can start to develop. Analyses of the
categorical perception effects in language learning models
have shown that language uses and modifies the space of
similarities between members of different perceptual and
linguistic categories (Cangelosi & Harnad, 2000).
To understand better the relations between
communication abilities and behavioral skills, the
correlations between fitness scores and a measure of the
quality of produced language have been computed. Figure 7
and 8 present the averages of the fitness curves, the
proportions of good speakers (i.e. language index), the
fitness/language correlation rall for the whole population,
and the fitness/language correlation rbest for the best 20
agents. Figure 7 refers to the 7 successful populations of the
From_50 - Parent speaker condition. Figure 8 refers to data
from the remaining 3 populations without communication.
For the computation of the language index based on the
proportion of good speakers, an agent is classified as good
speaker when it produces two opposite signals respectively
for the two objects in at least half of the 36 epochs. The
Pearson r correlations index was used.
Overall, the two figures show that the correlation
between the fitness of all 80 agents and their language
production index is positive and quite high (rall ≈ 0.5) after
good communication emerges. This can explain the
maintenance of good communication, since it reflects a link
between good speaking abilities and good comprehension
(i.e. behavioral fitness).
6
1
0.8
0.6
0.4
0.2
0
51
101
-0.2
0.4
Fitness(all)
r(best)
N. speakers
r(all)
0.2
0
51
61
-0.2
Figure 7 – Fitness curve of the whole population, number of
good speakers, fitness/language correlation rall for the whole
population and fitness/language correlation rbest for the best 20
agents (i.e. future parents and speakers). Average curves over
the 7 successful From_50-Parent speaker populations. Only the
data for generations 51-150 are shown. (see text for discussion)
1
0.8
0.6
0.4
0.2
There is no correlation between the fitness of the 20 best
performing agents and their speaking ability. The rbest stays
around 0 (with peaks of ±0.2) in both groups, for the
majority of generations. However, this correlation differs
significantly in the initial generations of the two groups of
populations with and without communication. At generation
51, there is a high positive fitness/language correlation for
the 20 best organisms (rbest=0.33) of the populations that
succeed. The correlation is much lower in the populations
that do not succeed (rbest=0.13). This indicates that, to
evolve communication, it is necessary to be in a population
where there is an initial positive correlation between the
fitness and language of the best performing individuals. This
initial correlation could be due to the role of hidden units,
where the linguistic production ability and the
comprehension/fitness abilities interact. As it has been
shown previously (Cangelosi & Parisi, 1998), the ability to
recognize and categorize the two objects can produce quite
distinct activation patterns in the hidden units. These will, in
turn, increase the possibility of initially producing different
signals for the two categories of objects.
The initial difference in the correlation rbest between the
successful and unsuccessful populations quickly disappears.
The correlation index becomes 0 after approximately 5
generations (cf. smaller chart of Figures 7 and 8, which
zoom in the first 10 generations). However, the initial
advantage of this high rbest correlation has the effect of
supporting the fitness/language correlation rall. In the
unsuccessful populations, this correlation goes down and
stays around 0 for all subsequent generations. Instead, in the
successful population, rall never reaches 0, and it starts to
grow since generation 4. The initial strengthening of the link
between production and fitness in the whole population will
subsequently help the establishing and maintenance of a
shared lexicon.
These correlation data helps answering question 4 and 5.
It explains the fine interrelationships between language
production, language comprehension, and fitness. In
addition, it highlights the role of cognitive factors in
supporting and favoring the emergence of communication.
Conclusions
0
51
101
-0.2
0.4
0.2
0
51
61
Fitness(all)
r(best)
N. speakers
r(all)
-0.2
Figure 8 – Data for the 3 unsuccessful From_50-Parent speaker
populations. (see text for discussion)
To summarize, the simulation of this evolutionary robotics
model of the evolution of communication shows that: (a) the
emergence of language brings direct benefits to the agents
and the population, in terms of increased fitness and
comprehension ability; (b) there is a benefit in
communicating with your kin-related agents (e.g. between
parents and children), since this improves the possibilities of
successfully evolving shared lexicons also by maintaining
stable and reliable signals; (c) good sensorimotor and
cognitive abilities permit the establishment of a link
between production and comprehension/behavioral abilities;
(d) the kinship relation between speaking parents and
listening offspring does not fully explain the emergence of
communication, since the rbest stays around 0 for most of the
7
generations – instead, this is important in the early stages of
communication because it exploits the cognitive benefits of
positive production/fitness correlations.
Most of these results have important implications for the
theories and hypotheses on the origins of language. For
example, this simulation highlights and explains the role of
cognitive factors in the emergence of communication
(Burling, 1993). In particular, the model supports the
hypothesis that the ability to form categories constitutes the
grounding for the subsequent evolution of words and
language (Harnad, 1996; Cangelosi & Harnad, 2000). In
addition, future developments of this model could also have
an impact on computational investigations of the mirror
neuron hypothesis for the origins of language (Arbib, 2002).
Further simulations will address in more detail the role
of sensorimotor coordination, cognitive, neural and social
factors in the emergence of complex communication
systems, such as syntactic languages. For example, the
authors plan to investigate (a) the factors that favor the
emergence of syntactic lexicon within such an evolutionary
robotics model, (b) whether listening to our own language
might contribute to the development of a communication
ability, and (c) whether language and communication might
lead to the development of internal categories that, aside
from communication, can be used by the robot to better
fulfill its own goals.
Acknowledgements
The work of D. Marocco and A. Cangelosi has been
partially supported by the UK EPSRC (Grant: GR/N01118).
References
Ackley D.H., Littman M.L. (1994). Altruism in the evolution of
communication. In: Brooks R, Maes P (eds) Proceedings of the
Fourth Artificial Life Workshop. MIT Press, Cambridge
Arbib M. (2002). Grounding the mirror system hypothesis for the
evolution of the language-ready brain. In Cangelosi & Parisi
(2002).
Baldassarre G., Nolfi S. and Parisi D. (2002). Evolving mobile
robots able to display collective behaviours, Technical Report.
Rome: Institute of Cognitive Science and Technologies, CNR.
Beer, R. (1995). A dynamical systems perspective on agentenvironment interaction. Artificial Intelligence, 72, 173-215
Braitenberg, V. (1984). Vehicles. Cambridge, MA: MIT Press
Brooks R.A. (1991b) New approaches to Robotics. Science,
253:1227-1232.
Cangelosi A. (2001). Evolution of communication and language
using signals, symbols and words. IEEE Transactions on
Evolutionary Computation, 5(2), 93-101
Cangelosi A., Harnad S. (2000). The adaptive advantage of
symbolic theft over sensorimotor toil: Grounding language in
perceptual categories. Evolution of Communication, 4, 117-142
Cangelosi A., Parisi D. (1998). The emergence of a "language" in
an evolving population of neural networks. Connection
Science, 10(2), 83-97
Cangelosi A., Parisi D. (2001). How nouns and verbs differentially
affect the behavior of artificial organisms. In J.D. Moore & K.
Stenning (Eds.), Proceedings of the 23rd Annual Conference of
the Cognitive Science Society, London: LEA, 170-175
Cangelosi A., Parisi D. (Eds.) (2002). Simulating the Evolution of
Language. London: Springer.
Floreano D., Urzelai J. (in press). Neural morphogenesis, synaptic
plasticity, and evolution. Theory in Biosciences.
Harnad S. (1990). The Symbol Grounding Problem. Physica D, 42:
335-346.
Harnad S. (1996). The origin of words: A psychophysical
hypothesis. In Durham W., Velichkovsky B. (Eds.)
Communicating Meaning: Evolution and Development of
Language. NJ: Erlbaum.
Hazlehurst B., Hutchins E. (1998). The emergence of propositions
from the co-ordination of talk and action in a shared world.
Language and Cognitive Processes, 13: 373-424
Kirby S. (2001). Spontaneous evolution of linguistic structure: An
iterated learning model of the emergence of regularity and
irregularity. IEEE Transactions on Evolutionary Computation
and Cognitive Science, 5(2): 102-110
Kirby S. (in press). Natural language from artificial life. Artificial
Life.
Lauria S., Bugmann G., Kyriacou T., Klein E. (2002). Mobile
robot programming using natural language. Robotics and
Autonomous Systems
Lipson H., Pollack J.B. (2000). Automatic design and manufacture
of robotic lifeforms. Nature, 406: (6799) 974-978.
Martinoli A. (1999). Swarm intelligence in autonomous Collective
robotics: from tools to the analysis and synthesis of distributed
control strategies. PhD Thesis. Lausanne - EPFL
Nolfi S., Floreano D. (1999). Learning and Evolution, Autonomous
Robots, 7(1): 89-113.
Nolfi S. (2000). Evorobot 1.1 User Manual. Technical Report,
Institute of Psychology, Rome
Nolfi S. (2002). Power and Limits of Reactive Agents.
Neurocomputing, 42:119-145.
Nolfi S., Floreano D. (1998). Co-evolving predator and prey
robots: Do 'arm races' arise in artificial evolution? Artificial
Life, 4(4), 311-335.
Nolfi S., Floreano D. (2002). Synthesis of autonomous robots
through artificial evolution, Trends in Cognitive Sciences 1,
31-37.
Nolfi S., Marocco D. (2002). Active perception: A sensorimotor
account of object categorization. From Animals to Animats 7.
Proceedings of the 7th Intl. Conf. on Simulation of Adaptive
Behavior.
Nolfi S., Floreano D. (2000). Evolutionary Robotics: The Biology,
Intelligence, and Technology of Self-Organizing Machines.
Cambridge, MA: MIT Press
Oliphant M. (1999). The learning barrier: Moving from innate to
learned systems of communication. Adaptive Behavior, 7(3/4).
Pfeifer R., Scheier C. (1999). Understanding Intelligence.
Cambridge, MA: MIT Press
Quinn M. (2001). Evolving communication without dedicated
communication channels. In J. Kelemen and P. Sosík (eds.),
Proceedings of ECAL01, Springer. pp. 357-366
Steels L. (1997) The synthetic modeling of language origins.
Evolution of Communication, 1(1).
Steels L. (1999). The Talking Heads Experiment (Volume 1: Words
and Meanings). Laboratorium: Antwerpen.
Steels L., Kaplan F. (2000). AIBO's first words: The social
learning of language and meaning. Evolution of
Communication, 4(1)
Steels L., Vogt P. (1997). Grounding adaptive language
games in robotic agents. In P. Husband & I. Harvey (eds).
Proceedings of the Fourth European Conference on Artificial
Life, London: MIT Press
8